Cryptocurrency risk assessment under a semi-nonparametric approach, risk measures and backtesting techniques

Resumen

The emergence of cryptocurrencies with a philosophy of independence from any institutions or central banks makes them a unique asset that has marked an important milestone in the financial and economic world. These assets also exhibit remarkable high volatility with frequent extreme values, which may cause instability in financial markets. Nor should we forget that cryptocurrencies are in their earliest stage and most probably its great volatility is due, in part, to this initial phase in which it finds itself. That is one of the purposes for carrying out a more comprehensive analysis on risk assessment of this industry. The focus of this research is illustrated through the following four chapters, based on modelling the full density of cryptocurrency returns with particular emphasis on providing accurate risk measures, i.e., fitting the tails of the distributions. With this aim, we analyze the conditional variance, modelled under GARCH-type models, considering the semi-nonparametric (SNP) approach based on Gram Charlier (GC) and Positive Gram Charlier (PGC) series. Firstly, this PhD Thesis explores the capacity of a method to approximate the cryptocurrency return conditional frequency distribution by endogenously selecting the best SNP expansion at any point in time compared to considering SNP expansions with a fixed length (number of parameters) in terms of the cumulative distribution function (cdf). The good performance of this new methodology compared to fixed-order GC expansion supports the thought about its usage in the future for forecasting risk. Following the line of a univariate perspective and considering SNP distributions as well as others parametric (Gaussian, Student’s t, Skewed-t), three different risk measures Value at Risk (VaR), Median Shortfall (MS) and Expected Shortfall (ES) have been assessed through backtesting techniques and a wide variety of tests. This comprehensive analysis for Bitcoin and five of the most representative altcoins shows that flexible SNP approaches outperform risk measures of most crypto assets (especially Bitcoin) and tend to provide the most conservative risk assessment. Furthermore, MS seems to be a robust-to-outliers and reliable risk measure for cryptocurrencies and discuss the choice of the appropriate probability levels according to the assumed distribution. The evidence supports that MS might be an accurate alternative to VaR and ES. Under a multivariate analysis, it is implemented a flexible and accurate new methodology for portfolio risk management that consists of computing pairwise conditional correlations under bivariate marginal SNP distributions with PGC through the Dynamic Conditional Correlation (DCC) model. This method tries to solve the ‘curse’ of dimensionality triggered in DCC models when there are large portfolios compared with a parsimonious model as Dynamic Equicorrelation (DECO). Both models are tested for a portfolio of three cryptocurrencies (Bitcoin, Litecoin and Ripple) through backtesting techniques for VaR, MS and ES. In the light of the results, both models show good performance but the new method might be an excellent proposal for analyzing accurately large portfolios. Finally, a summarize of all the conclusions are presented at the end of this PhD Thesis as well as some recommendations for future research.